Skip to content

intel/fs: Use nir_opt_uniform_atomics()

Kenneth Graunke requested to merge kwg/mesa:brw-diverge into main

(first two patches are from Caio's other MR, !15400 (merged))

This series enables nir_opt_uniform_atomics() to significantly improve performance of atomics to a uniform address. It improves performance of one benchmark by 30%, and Shadow of the Tomb Raider by 5.5%. More details in the commit messages.

This is the first use of convergence information by the Intel backend, so we enable that as well.

This is currently limited to Broadwell and later because there are some Vulkan CTS and crucible test failures on Haswell that I haven't managed to figure out yet. I suspect there is just a bug and we could enable it in a follow-up series at some point.

Merge request reports